Feature Value and Overview
Nosto's Hybrid Vector Search is an advanced search technology that merges traditional keyword matching with vector-based semantic understanding. This intelligent approach goes beyond simple character matching to understand the intent of a user's query, leading to a smarter shopping experience.
What it Does
The system automatically decides whether to use only traditional keyword search results or blend them with vector search results based on specific scenarios that can be individually configured and controlled by the user.
Vector search works by capturing the conceptual similarity between the user's search query and your product data, ensuring that conceptually related products are found even without exact keyword matches.
User Configuration and Control
You can actively manage how Hybrid Vector Search works through the Nosto Admin UI:
Basic Activation: Toggle to activate the vector search component of hybrid vector search.
Vector Search Minimum Similarity: Determines if vector search returns only strongly similar products or also more loose matches.
Trigger Scenarios: The system is designed to trigger hybrid search for 0-Results Pages and for Low CTR/CR Queries. You can configure the performance thresholds that define a 'low' performing query. The additional information when choosing a certain value shows you the expected ratio of hybrid search based results. This helps you understand the impact of the chosen configuration.
Keep in mind that hybrid vector search is designed to expand keyword search for queries where keyword search is poorly performing. Hybrid search is not meant to fully substitute keyword search - keyword search has it's own strengths (indicated by the high performing queries). You can check search analytics to get a better feeling for the shop’s average CR/CTR and select a value that's indicating a poor CTR or CR based on this.
Query Management: You have granular control to either exclude specific queries from ever including vector-based results or to force hybrid vector search activation for queries that the automatic system doesn’t cover.
Score Insights: Similarly to keyword search, you can see the impact ratio between vector search and merchandising rules in the score insights UI. For hybrid results, you can see the ratio between keyword search, vector search and merchandising impact.
Merchandising: All your existing merchandising rules, such as pinning and boosting, remain fully applicable to results surfaced via vector search.
Searchable Fields: To achieve the best results, ensure that only fields containing precise and concise product information are selected as searchable fields. This is as important for keyword search as it is for vector search.
Integration: Activation is done solely through the Admin UI. For front-end customisation, the Search API response will include new information to identify which results were powered by vector search. For more information on the API, see the related Tech Docs.
Data Vectorisation Frequency: Vectors are generated every 6 hours.
Analytics
All accounts using hybrid vector search, will see enhanced search analytics that highlight the impact of vector search to the overall search metrics. The highlighted percentage in the blue box shows the percentage of all vector search interactions that contributed to a certain metric, available for:
Total searches
Orders
Sales
Product clicks
Moreover, there are also some vector search specific metrics available in the search analytics tab:
Amount of vector search responses
Amount of avoided no results pages
If hybrid vector search is enabled for your account, there are two new tabs added to the search queries table:
Only Vector Results: Showing all queries without any keyword search results but with vector search results
Hybrid Results: Showing all queries with combined keyword and vector search result sets
If you are not using NostoJS' automatic tracking, you need to update your tracking implementation to make sure, vector search responses are tracked as such. See the Tech Docs for detailed guidance.